Trained on Krea 2 Raw using a curated dataset of 22 images. This LoKr adds only marginal value (slightly more vibrant colors and less clutter). Krea 2 Turbo already generates convincing Kinkade-style images out of the box. Here's a side-by-side comparison:<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/31d976fd-c790-4aa5-a324-36304c90ea69/original=true/31d976fd-c790-4aa5-a324-36304c90ea69.jpeg" />Recommended settings:<ul><li>8 steps</li><li>cfg 1.0</li><li>sampler: <code>er_sde</code></li><li>scheduler: <code>simple</code></li></ul>Prompting:Trigger word is not required but use <code>Thomas Kinkade style</code> or similar wording for best results. For typical Kinkade scenery (cabin, village, snow, etc), short prompts in the form of comma-separated phrases will suffice. But for non-typical use, long prompts produce much better results.Known issue:Outputs may contain signatures when using long, natural-language prompts mentioning art style and medium. This comes from the base model's pretraining, not from the LoKr or dataset cleanup failures.I personally use <a rel="ugc" href="https://huggingface.co/Comfy-Org/Mage-Flow">Mage Flow Edit</a> to remove the signatures. Check out the second showcase image for an example workflow.

Files with SHA256: b688553255b62ae7df57086a1ad63b2c9d7cdbce9c615c16b036fb16ee33ef2d